Generative Models Show Persistent Bias Toward Traditional Attire

A new paper by Abdulkareem Alsudais published on September 30, 2026 examines how popular text‑to‑image models depict individuals from 206 nationalities when prompted to illustrate everyday activities ¹.

The authors generated 2,060 images across five activity scenarios using prompts that specified nationality names for each model—DALL‑E 3 and Gemini 3 Pro Preview—and evaluated them for cultural representation ¹.

Results show that 28.4% of the images depicted individuals wearing traditional attire, including outfits impractical for the requested activity in several cases ¹. The pattern was statistically significant and associated with geographic regions; Middle East & North Africa were disproportionately affected, while World Bank income groups also correlated with higher rates of traditional depiction ¹. Similar region- and income-linked patterns appeared for images labeled as featuring impractical attire in two athletics-related activities ¹.

To assess image–text alignment, the team used CLIP, ALIGN, and GPT‑4.1 mini and scored 9,270 image–prompt pairs. Images labeled as featuring traditional attire received significantly higher alignment scores when prompts included country names, indicating a stronger perceived match between prompt content and visual output ¹.

Expert commentary notes that CLIP embeddings themselves can introduce bias if not carefully considered—previous studies showed that even small changes in prompt wording significantly impact alignment scores across different demographic groups ¹.

DALL‑E 3’s larger training dataset (~2 million images) yields some improvement over earlier models; Gemini 3 Pro Preview still struggles substantially when prompted about less represented nationalities such as Bhutanese people engaging in non-traditional occupations ¹.

These findings reinforce concerns raised last year regarding algorithmic amplification of stereotypes within generative AI systems ¹.

The study highlights how current text-to-image generators may unintentionally perpetuate cultural biases through their training data and prompting mechanisms. While technical advances have expanded model capabilities, this research underscores the need for more inclusive datasets and evaluation metrics that account for regional diversity and socioeconomic context. The results suggest that developers should incorporate explicit checks against stereotypical depictions during both training and deployment phases to mitigate unintended reinforcement of cultural assumptions.